Wijenayake-S/dr-severity-efficientnetb0-cbam

Diabetic Retinopathy Severity Grading (EfficientNetB0 + CBAM)

Five-class ETDRS severity grading from retinal fundus photographs.

Author: H.W.P.S.M. Wijenayake Project: NeuroGSD - AI Based Early Detection System for Common Eye Diseases Institution: Faculty of Information Technology, University of Moratuwa, Sri Lanka Supervisor: Dr. L. Ranathunga Domain expert: Dr. K. M. P. K. Bandara

Classes

0-No DR | 1-Mild NPDR | 2-Moderate NPDR | 3-Severe NPDR | 4-PDR

Architecture

EfficientNetB0 (ImageNet) + CBAM channel/spatial attention + global average pooling + dropout(0.4) + softmax. Two-stage transfer learning: frozen backbone first, then fine-tuning from block5a with BatchNorm layers held frozen.

Files

File	Input size	Gaussian sigma
dr_448.keras	448x448	20
dr_224.keras	224x224	10

config.json holds the full preprocessing specification. Preprocessing must match exactly - the model expects Ben Graham enhanced images, not raw fundus photographs.

Preprocessing

1. Crop black border (intensity threshold 7)
2. Resize to the model's input size
3. Ben Graham enhancement: addWeighted(img, 4, GaussianBlur(img, (0,0), sigma), -4, 128)
4. Rescale by 1/255

Results (held-out test set, n=262)

Configuration	Accuracy	QWK	Macro-F1
224px argmax	0.6718	0.6209	0.4783
448px argmax	0.6985	0.7420	0.5700
448px + TTA (deployed)	0.7176	0.7695	0.5722
448px + 224px ensemble	0.7290	0.7848	0.5775

Per-class recall for the deployed configuration: No DR 0.876, Mild NPDR 0.122, Moderate NPDR 0.750, Severe NPDR 0.727, PDR 0.400.

Referable DR (classes 2-4 vs 0-1), 448px argmax configuration: AUC 0.942, sensitivity 0.838, specificity 0.923.

91.2% of predictions fall within one severity grade; 0.0% are misgraded by three or more.

Training data

1,744 fundus images at 512x512 native resolution, stratified 70/15/15 split (1220 / 262 / 262). Class counts: No DR 1017, Mild NPDR 270, Moderate NPDR 347, Severe NPDR 75, PDR 35. Imbalance ratio 29:1, addressed with square-root-smoothed class weights.

Limitations

• Mild NPDR recall is 0.39.
• PDR results are unreliable (n=4).
• Severe NPDR (n=11) is also low-support; its 0.727 recall is indicative only.
• Single-source dataset; generalisation across cameras and populations is untested.
• Severity labels were graded by H.W.P.S.M. Wijenayake under the domain expert's guidance and subsequently reviewed and confirmed by him - not independently double-graded.

Statistical caveats

With 262 test images the 95% confidence interval on accuracy is approximately +/-5.5 percentage points. Stratified 5-fold cross-validation over all 1,744 images is planned to give more reliable per-class estimates.

The 448px + TTA configuration was deployed rather than the marginally higher-scoring ensemble: the 1.1-point accuracy difference amounts to 3 images and falls within the confidence interval, while the ensemble doubles inference cost and requires two parallel preprocessing pipelines.

Intended use

Research and educational use only. This model is not a medical device and must not be used for clinical diagnosis. Any screening deployment would require prospective clinical validation and regulatory approval.